Starting from July 2026, the gIB NCCL plugins are integrated directly into NVIDIA's NGC deep learning containers, providing an "out-of-the-box" experience that lets you achieve expected performance for collective operations on Google Cloud GPU platforms without manual configuration or tuning.
Overview of the integration
The integrated NGC base images include the following artifacts:
- Environment plugin: Automatically verifies Google Cloud residency and detects the specific hardware profile (such as H200 or GB200) to dynamically apply the optimal environment variables at runtime.
- Network plugin (gIB): Provides guest-level multi-pathing for high-performance point-to-point transfers and improved load balancing on Google's networks.
- Tuner plugin: Dynamically optimizes collective communication operations specifically for Google Cloud hardware topologies at scale.
- Profiler plugin (CoMMA): Features advanced telemetry and profiling to detect when a workload isn't progressing and identify stalled NCCL processes.
The plugins detect when they're running on supported Google Cloud platforms and automatically self-configure, while gracefully falling back to standard configurations on unsupported environments.
Benefits
By using the NVIDIA NGC containers with the embedded Google Cloud plugins, you can expect the following:
- Seamless delivery: Immediate performance on all supported Google Cloud GPU platforms without complex setup.
- Optimal performance: Automated, internal environment detection ensures optimal collective bandwidth and latency out-of-the-box.
- Diagnostics ready: Advanced telemetry client built-in, helping Google support to better assist you.
Supported platforms
The NGC containers with gIB NCCL plugins support the following Google Cloud machine types:
- A3 Ultra (H200)
- A4 (B200)
- A4X (GB200)
- A4X Max (GB300)